Performance on Activities of Daily Living and User Experience When Using Artificial Intelligence by Individuals With Vision Impairment

Abstract

Purpose: This study assessed objective performance, usability, and acceptance of artificial intelligence (AI) by people with vision impairment. The goal was to provide evidence-based data to enhance technology selection for people with vision loss (PVL) based on their loss and needs.

Methods: Using a cross-sectional, counterbalanced, cross-over study involving 25 PVL, we compared performance using two smart glasses (OrCam and Envision Glasses) and two AI apps (Seeing AI and Google Lookout). We refer to these as assistive artificial intelligence implementations (AAIIs). Completion and timing were quantified for three task categories: text, text in columns, and searching and identifying. Usability was evaluated with the System Usability Scale (SUS).

Results: The odds ratios (ORs) of being able to complete Text tasks were significantly higher when using AAIIs compared to the baseline. OR when performing “Searching and Identifying” tasks varied among AAIIs, with Seeing AI and Envision improving the performance of more tasks than Lookout or OrCam. Participants expressed high satisfaction with the AAIIs.

Conclusions: Despite the findings that performance on some tasks and when using some AAIIs did not result in a greater number of PVL being able to complete the tasks, there was overall high satisfaction, reflecting an acceptance of AI as an assistive technology and the promise of this developing technology.

Translational Relevance: This evidence-based performance data provide guidelines for clinicians when recommending an AAII to PVL.

Introduction

Whereas past assistive technologies (ATs) have provided reading support for people with vision loss (PVL) by manipulating the appearance of text on screens, recent software and hardware advances, innovations in computer vision, and the development of algorithms commonly known as artificial intelligence (AI) offer promising solutions. AI simulates human intelligence processes, such as learning, reasoning, visual perception, speech recognition, decision making, and language comprehension. The growing ubiquity of AI has significantly increased access to information for nearly everyone, including PVL. Moreover, deep-learning approaches to AI allow for rapid retrieval and processing of information that support processes, such as reading text, recognizing objects and persons, and describing scenes. With increased computing power and improvements in digital cameras, integrating AI into portable applications, such as “smart glasses” and cellular phone apps, holds great promise for the future of AT, offering hope for a more inclusive and accessible future. 

Recently, there has been considerable interest in the engineering literature regarding applying AI to the needs of PVL. A review of this literature in 2023 found 646 published studies on vision assistive AI in just the previous 2 and a half years.¹ Diverse and innovative applications of AI in developing AT for PVL share the goal of enhancing independence and quality of life.^2–9 Walle et al.⁸ and others have outlined the issues to be resolved in the use of AI for PVL, including the types of information to be obtained (e.g. object recognition, navigation, and scene description), the location of acquisition hardware (e.g. wearable, cane-mounted, and mobile phones), the types of sensors (e.g. cameras, global positioning system (GPS), position sensors, and infrared detectors), the algorithm used (e.g. YOLO, SLAM, and Kalman filters), the limits imposed by the processing load (to be done locally or remotely), the nature of the user interface (e.g. buttons, gestures, and voice), and the choice of feedback modalities (e.g. auditory and tactile). Despite these multiple obstacles, proposed solutions have been appearing rapidly. Coupling innovative AI technologies with the capabilities of wearable devices and mobile phone apps can provide users with tools for navigating and understanding the visual world. The potential of AI to transform the lives of PVL has generated enthusiasm in the vision-loss community. However, the functionality of AI in the hands of PVL must be objectively quantified. In this study, we examined how these advanced technologies perform, given the constraints of vision loss. 

Methods

Study Design

A cross-sectional, counterbalanced, cross-over study compared performance using four assistive artificial intelligence implementations (AAIIs): two smart glasses OrCam MyEye 2 Pro, version 9.21 (https://www.orcam.com/en-us/home, $4250 as of July 31, 2024), and Envision Glasses Home Edition (https://www.letsenvision.com/glasses/home, $2499 as of July 31, 2024); and two mobile phone applications Seeing AI version 1.0.1 on an iPhone 13 pro, IOS 17.6 (https://www.seeingAI.com, $0), and Lookout by Google for Android version 4.3 on a Pixel 4, Android version 11 (https://www.lookout.com, $0). All AAIIs provided audio feedback, such as describing scenes and reporting the presence and descriptions of people and objects, to allow users access to text and information about their surroundings. 

Recruitment and Inclusion Criteria

Participants were recruited from the Lighthouse Guild's low-vision clinic and rehabilitation services. We asked the staff to refer patients for screening and advertised on the research page of the Lighthouse Guild's website. Potential participants were screened based on the inclusion criteria, including being at least 18 years old, visual acuity from 20/100 (0.7 logMAR) to no light perception, and a cognitive function score of ≥10 (based on the 6-item Cognitive Impairment Test).¹⁰ Additionally, individual goals were measured with the Participatory Activity Inventory (PAI),¹¹ and technology goals were determined using the Matching Person and Technology (MPT) assessment.¹² Participants were enrolled based on their willingness to explore the use of AT to achieve their goals and commitment to attend the clinic for five sessions. Participants were excluded if they had prior proficiency with any AAII to be tested in the study. 

Study Procedure

Data were collected during five sessions – baseline performance with no AI and wearing usual refractions (session 1) and four subsequent visits, each with one of the four AAIIs (sessions 2–5). Sessions ran for approximately 90 minutes, including breaks, and were scheduled about a week apart. The order of AAII testing was counterbalanced across participants. Users practiced each AAII's range of functionality by performing training tasks. We encouraged participants to become familiar with the physical device, open the software, scroll among options, and learn appropriate camera framing techniques for each task. The practice did not use the versions of the tasks tested in the data collection part of the study, and performance on the practice tasks was not scored. 

Before the testing phase began, the researchers quantified the performance of all settings of a given AAII for each task. They determined the setting that yielded the best results for a task and AAII (Table 1). During testing, the experimenter launched the app and selected the appropriate setting for the task. 

Table 1.

 

The Best Settings for the Tasks Determined Prior to Testing

 

Outcome Measures

We collected descriptive data at baseline on eye disease/cause of vision loss, visual acuity measured with an Early Treatment Diabetic Retinopathy Study (ETDRS) letter chart at one meter (Precision Vision ETDRS Chart No. 2112, https://precision-vision.com), onset of vision impairment, self-reported severity of visual impairment, age, sex, race, comorbid conditions, education, living situation, employment, participants’ needs and goals (measured with the PAI), and familiarity with and receptivity to AT to achieve their goals (measured with the MPT). 

Performance and timing on 14 vision-related activities of daily living (ADL) were evaluated at each session. The tasks were carefully chosen based on published literature, patients’ needs, and the capabilities of the AAII.^13–20 The tasks were divided into three categories based on the primary visual requirement of the tasks. All tasks were timed and scored except for reading a newspaper article, which was just scored. The criteria for scoring accuracy of an AAII when used by PVL and participants’ successful task completion when using an AAII are shown in Table 2. 

Table 2.

 

Scoring Criteria for the Tasks

 

Category 1, “Text,” included tasks typical of those evaluated in vision-related ADL questionnaires. The tasks were: (1) read a newspaper article (from The New York Times Large Print Weekly, text size 2 M) and describe the article’s contents; (2) read an invoice and report the amount due (text size 2 M); (3) read a handwritten five-item grocery list and report the non-food item (text size equivalent 2 M); (4) report the name of the medication and dose instructions on a pill bottle (text size for medication 2.75 M and instructions 2 M); (5) identify a banknote; and (6) read a street sign at eight meters (text 74 M). 

Category 2, “Text in Columns,” included (7) read a table of contents from the newspaper and report the page number of a given article (e.g., on which page is the Business Section? text size 4 M); and (8) read a digital TV channel guide from Spectrum presented on a 24 inch TV at 2 meters (equivalent text size 3 M). The task was to report the channel for a given show (e.g. On which channel is Major League Baseball [MLB]?). 

Category 3, “Searching and Identifying,” included (9) find the matching sock. A sock of a given color was shown, and the task was to find that color sock in a group of four socks of different colors; (10) scan a grocery product box to identify the product from the barcode; (11) match a photograph of a person. One photograph was shown initially, and the task was to match the same photograph in a group of three photographs; (12) describe a scene. Participants stood at 1.5 meters in front of a color poster of a landscape so that the image filled the camera's field of view; (13) find a person in a room. A person stood quietly at one of five positions in the testing room; and (14) identify a room when standing at the entrance. The rooms were an office, living room, bathroom, kitchen, and store. 

To prevent learning effects, a different version of each task was used in each session (e.g. 5 different newspaper articles, 5 different medicine bottles, and 5 different street signs), and the order of the task versions was counterbalanced across AAIIs. 

At the end of each session, the participant completed two surveys: (1) a subsection of the MPT (form 4) titled Assistive Technology Predisposition Assessment. The questions were: “This AT will help me achieve my goals,” “This AT will fit well with my accustomed routines,” “This AT will physically fit in all desired environments,” and “I will feel comfortable using this AT around others”); and (2) the System Usability Scale (SUS), which has 10 items scored on a 5-point Likert scale and has often been used in studies on AT due to its good psychometric properties.^21,22 After each session, participants completed free-form responses to the question, “For which tasks would you use this AT, including any tasks in addition to those tested today?” 

Statistical Analyses

Log odds ratios (ORs) with Firth Bias were calculated, indicating the odds of a person being able to perform the task when using an AAII compared with the odds of performing the task at baseline (without any AAII). Person abilities were compared with AAII abilities descriptively. To detect overall differences in timing among conditions (baseline and the 4 AAIIs) across each of the multiple tasks, nonparametric Friedman One-Way Repeated Measure Analyses of Variance by Ranks were performed (the data failed to meet the assumptions of parametric testing). A Benjamini-Hochberg correction was used to control the false discovery rate due to multiple Friedman tests.²³ Post hoc (pairwise) Wilcoxon signed-rank tests were used to determine whether there were statistically significant differences among conditions. Multivariate logistic regression analyses were used to identify baseline performance and outcomes predictors. All analyses were performed in R Studio version 4.2.1, and a P value of ≤ 0.05 was used to determine statistical significance. 

Results

Participant Characteristics

The Institutional Review Board of the Lighthouse Guild in New York, NY, approved the study. Twenty-five eligible participants gave informed consent and completed the study. They were compensated for their participation. The average age of the participants was 64 years (range = 30–83 years), visual acuity ranged from 0.7 logMAR to no light perception, and text-to-speech was the most used AT at baseline (84% of participants; Table 3). Fifty-two percent were White, 40% were Black, and 8% were Asian. In response to the MPT questions about previous experience with any assistive technology, 62% were satisfied, 28% were neutral, and 10% were unsatisfied. 

Table 3.

 

Participant Characteristics (n = 25)

 

Outcomes

Initially, the data were examined for order effects because there might have been a practice effect over the AAII conditions. There was no significant effect of order for any tasks (Friedman One-Way Repeated Measure Analyses of Variance by Ranks), confirming that the counterbalancing was effective. 

We compared the ORs for the number of participants who could complete the tasks using an AAII (“N able” in Table 4) relative to those who could do the tasks at baseline. We found that the ORs of PVL being able to complete tasks in the text category were significantly higher for 100% of the tasks when using Envision, Seeing AI, and Lookout and for 83% of the tasks when using OrCam (see Table 4). In the text in columns category, significantly more participants succeeded in finding a page number and a channel for a TV show using OrCam (100%) and Seeing AI (50%). In contrast, using the other two AAIIs did not significantly increase the ORs of completing the text in columns tasks, predominately because they read the text column-wise. For the non-text searching and identifying tasks, using Seeing AI resulted in higher ORs of being able (83%), followed by OrCam (50%), Envision (50%), and Lookout (33%). In Table 4, ORs were indicated as significantly higher in bolded black text (ɫ) and significantly lower in bolded red text (‡). Nonsignificant findings are not bolded. 

Table 4.

 

Odds Ratios of PVL Being Able to Perform a Task

 

Timing

The time to complete the task was measured as the total time it took the AAII to capture, process, and announce the information, plus the time it took the participant to interpret and report the information provided by the AAII. The overall Friedman analyses on the timing data for tasks in the text category showed significant time differences across AAII conditions for the invoice (χ² = 34.3, P < 0.001), handwriting tasks (χ² = 75.2, P < 0.001), medicine bottle (χ² = 34.0, P < 0.001), banknote (χ² = 45.3, P < 0.001), and street sign (χ² = 43.1, P < 0.001) tasks. Timing was not measured for the article task due to the different text lengths of the articles. In the text in columns category, neither the table of contents nor the TV guide task showed significant differences across testing conditions. For tasks in the searching and identifying category, the following tasks showed significant differences across testing conditions: color matching (χ² = 60.1, P < 0.001), face matching (χ² = 31.4, P < 0.001), landscape (χ² = 83.4, P < 0.001), and room identification (χ² = 9.3, P < 0.001). The overall Friedman test outcomes were not significant for timing for the barcode task and the find a person task. 

Post hoc Wilcoxon signed-rank analyses for the tasks found to have significant main effects are shown in Table 5. For the invoice and the medicine bottle tasks, timing was significantly faster than baseline when using all AAIIs. Timing for the handwriting task was significantly faster than baseline when using Envision, Seeing AI, and Lookout, but not OrCam. For the banknote task, only Seeing AI and Lookout were significantly faster than the baseline. All but OrCam allowed significantly faster than baseline performance on the street sign task. When doing tasks in the searching and identifying category, using OrCam and Lookout resulted in equivalent or statistically slower performance than baseline, and using Seeing AI allowed equivalent or significantly faster times for the landscape and room tasks. 

Table 5.

 

Wilcoxon Signed-Rank Test Results Comparing the Performance of an AAII to Baseline, Shown Only When the Overall Friedman Test was Statistically Significant (n = 25, P ≤ 0.05)

 

Comparison of AAII's Accuracy and Participants’ Task Success

We collected data on AAII's accuracy and participants’ task success. One might expect the relationship between AAII accuracy and participants’ task success to be unitary; however, this was not always the case. In Figure 1, the dots are placed at the X-values for the percentage of tasks the AAII could do correctly (based on the criteria described in Table 2). The arrows represent the magnitude of additional tasks completed by the participants. For example, the OrCam device accurately completed an average of 55% of the text tasks; however, the participants completed 33% more. Similarly, participants could complete 10% more tasks than Envision, 5% more than Seeing AI, and 26% more than Lookout. This better performance reflects the failure of the AAII to read correctly, including missing words and phrases, mispronounced words, misplacement of line orders, etc., plus the added effects of human intelligence using context. In these cases, participants accurately inferred information about the text that the AAII did not specifically report. In contrast, no arrows extend from the dots for tasks in the searching and identifying category where language context did not play a role in task completion. That is, participants were only successful to the extent that the AAII reported the information required to complete the tasks. 

Figure 1.

 

Comparison of AAII success versus participant success.

Predictors of Outcomes

We explored the predictive value of sex, age, visual acuity, diagnosis, and responses to the baseline PAI questions about perceived performance on similar tasks to our outcome measures. Using Multivariate Logistic Regression analyses, we found that being able to do the tasks at baseline was significantly predicted by logMAR visual acuity and lower PAI-measured perceived difficulty of doing the same tasks (Table 6). 

Table 6.

 

P Values Associated With Predicting Performance at Baseline

 

Table 7.

 

P Values Associated With Predicting Change in Performance

 

We also explored the predictive utility of demographic and perceived task difficulty in terms of changes in task performance with each AAII compared to baseline (improved, unchanged, or worsened). For all task categories, higher perceived difficulty significantly predicted improved performance with an AAII (Table 7); that is, more participants who reported higher perceived difficulty on the PAI administered at baseline had improved task performance when using an AAII than those reporting lower difficulty. In addition, more participants with poorer baseline visual acuity had improved task performance when using an AAII than those with better acuity. 

Graphical Representation

As we have done previously,²⁴ symbols were used to interpret the statistical findings from Tables 2 and 3 visually. Figure 2 aims to help practitioners recommend AT to PVL by offering a framework based on the AAII’s performance. The tasks represent the goals and needs of the patient, and the symbols represent the AAII’s performance. A large dot indicates where using an AAII resulted in a significantly higher OR, and the Wilcoxon signed-rank tests of quantitative timing data showed significantly faster performance than baseline. A small dot indicates that there was only a significantly higher OR, whereas the Wilcoxon signed-rank tests showed no statistically significant difference in timing relative to the baseline. An equal sign indicates when the AAII’s performance, as measured by both statistics, was not significantly different from the baseline. A small “X” shows a significantly lower OR whereas the Wilcoxon signed-rank tests showed no statistically significant difference, and a large “X” is drawn where both showed statistically worse performance than baseline. Across the tasks, using Seeing AI resulted in the best overall performance. 

Figure 2.

 

Graphical representation of devices based on ORs and the Post hoc Wilcoxon Signed-Rank findings.

Satisfaction

Based on the MPT (form 4), participants reported that all the AAIIs mostly met their desired goals, routines, and environments and that they did not feel self-conscious using them, with a median Likert score of 5 out of 5 for Seeing AI, 4 out of 5 for OrCam, 3 for Envision, and 3 for Lookout. 

SUS data were transformed and calculated according to the standard method.²¹ The mean score across participants was 77.9 for Seeing AI, 62.5 for Envision, and 57.2 for OrCam and Lookout (higher scores represent higher perceived usability). Person measures were calculated using Rasch analysis (Winsteps) for each AAII SUS score and compared using a One-way Repeated Measures Analysis of Variance. There was no significant effect of the AAIIs on perceived device acceptance (F = 1.04, P = 0.38). 

Perceived Use

At the end of each testing session, participants were asked to list the tasks they would use that AAII for. Their answers were binned into categories: reading, scene description, shopping, money identification, etc. There were no significant differences in the distribution of tasks across the AAIIs, and the total occurrence for each category is plotted in Figure 3. By far, the most often reported perceived use was for reading tasks. 

Figure 3.

 

Distribution of perceived use for AAIIs.

Discussion

Recent strides in machine learning have paved the way for the development of AI technologies that can perform tasks traditionally requiring human intelligence. Computational Neural Network innovations have revolutionized computer vision, leading to considerable progress in image recognition, object detection, and other visual tasks. This, combined with sophisticated Large-Language models that process and generate text and capture the intricacies of language, has supported translation into real-world applications that directly assist PVL (such as Seeing AI and Lookout). The future looks promising, as modern computer methods are increasingly being harnessed for assistance to PVL, as evidenced in recent review articles.^1,25–27 

In the present study, we compared the success and ease of use of four AAIIs to a baseline condition with no AAII by collecting objective performance data on real-world tasks and subjective data on usability and acceptance. Our findings showed that the four AAIIs we evaluated were most effective when used for tasks in the text category. The ORs of completing tasks were significantly higher than at baseline for five of the six tasks in the text category when using OrCam and for all six tasks using Envision, Seeing AI, and Lookout. For the searching and identifying category, ORs of being able to complete tasks were significantly higher than at baseline for five of the six tasks when using Seeing AI, four tasks with Envision, three with OrCam, and only two with Lookout. 

Previous studies have explored the functionality of AI with PVL. Granquist et al.²⁸ compared performance using the OrCam MyEye 1 to an earlier version of Seeing AI using six reading tasks. With a sample size of 7, they found that 71% of the tasks could be completed using OrCam, and 55% could be done using Seeing AI. This difference in performance was not significantly different. Likewise, participants' ease of use ratings were not significantly different between the two AAIIs. In contrast, we found that using Seeing AI allowed more participants to perform reading tasks faster, perhaps due to differences in the newer versions of each AAII. Nguyen et al.²⁹ studied the effectiveness of OrCam MyEye 2 in 20 patients with inherited retinal disease. Participants were trained and provided with an OrCam for 5 weeks of home use. At follow-up, there were significant improvements in self-reported functioning on the National Eye Institute Visual Function Questionnaire (NEI-VFQ)’s visual functioning subscale, the reading goal of the PAI, and persons’ scores on the OrCam Function Questionnaire (OFQ). There was no comparison with other AAIIs in this study, and the amount of at-home usage was not quantified. Amore et al.³⁰ examined the efficacy of the OrCam MyEye 2 using reading, money, color, face, and object recognition tasks. Reading tasks improved the most, with an average of 71% of the participants, but other tasks improved by only 26% (from their table 1). Bhagat et al.³¹ assessed accuracy and usability in an unstated number of PLV users. They reported 100% accuracy for printed text for Seeing AI, Envision, and Lookout, 90% accuracy for handwritten paragraphs for Seeing AI and Lookout, and good usability for Seeing AI and Lookout, with medium usability for Envision. Our findings are generally consistent with these past reports. 

We expected that participants with eye diseases characterized by central vision loss would benefit most from an AAII. However, diagnosis, along with sex and age, did not predict the performance at baseline or outcomes in our current study. Although we found no statistical relationship between clinical and demographic variables and outcomes, perceived task difficulty assessed with the PAI at baseline significantly predicted the objective performance of the same tasks at baseline. Similarly, Szlyk et al.³² reported an average correlation between perceived and functional performance of 0.50 ± 0.12. Latham and Usherwood¹⁶ reported an average correlation of 0.61 ± 0.15 between self-reported difficulty and rate of task completion, and Kartha et al.³³ found a correlation of 0.60 between self-reported visual ability and actual task performance. In our current study, perceived ability, as recorded by PAI questions, significantly predicted performance at baseline. This finding adds to the external validity of some of the questions in the PAI. 

We assessed usability with the SUS, an instrument sensitive to differences among products and interfaces.^21,22 Amore et al.³⁰ reported that the SUS score for OrCam was 63.5, but their study design did not compare the AAIIs. The usability of the AAII we tested ranged from 77.9 for Seeing AI to a low of 57.2 for both OrCam and Lookout. 

We found that the most frequently perceived potential use of an AAII was for reading (47.4%). This observation aligns with previous reports showing that reading improvement is the most common goal of PVL when seeking rehabilitation.^34–36 and as a use for technology.¹ In the past, this stated goal of improving reading had been interpreted as a need for training to allow patients to best use their residual vision for reading³⁷ and for technologies that alter text through manipulation of luminance, contrast, and magnification.^38–45 The data in Figure 3 suggest another interpretation; that is, PVL wish to access the content of text, even if the visual system does not process the text. They overwhelmingly stated that they would use the AAII for reading, even though the technology did not always “read” in a manner typical of visual reading. That is, audio reading is a very acceptable way to access text for PVL. 

A parallel exists between the AAII findings and our previous work using head-mounted displays.²⁴ Although image-manipulation approaches to AT significantly increased the number of reading tasks PVL could do, they provided little assistance with other ADL tasks. Understanding written material is a straightforward task as all the information is included in the text, and the function of technology is to present that information in an understandable format. On the other hand, the information needed to complete an ADL task is not always lexical, and inferences need to be made based on a separate set of variables for each task. Current technologies, including AI, do not gather information top-down and fail to assist in tasks requiring active perception. 

Strengths and Limitations

Our experimental design and choice of tasks and outcome measures allow for multiple AAIIs to be compared in small-scale studies that provide objective and subjective information, which is imperative for evaluating quickly advancing technologies. In addition, we used a cross-over design and counterbalanced the order of the tasks to guard against order and sequence effects. We included a heterogeneous group of participants, varying in age, sex, onset and severity of the visual impairment, and various eye conditions (including glaucoma, macular degeneration, cone dystrophy, and retinitis pigmentosa). This diversity contributes to the richness of our data and aligns with claims that the AAIIs are suitable for a large variety of PVL. Moreover, our study was conducted without the influence of commercial parties that could compromise the objectivity and integrity of the results. 

Participants used the AAII after training and instructions. To solely test the AAII’s accuracy and task completion, we effectively made the AAII interface fully accessible by determining the best setting for each task for each AAII before data collection began (see Table 1) and then having the experimenter launch the app and select the ideal setting for each task. It is unknown whether the patterns of results would have been different if participants had had the opportunity to use the AAII at home for an extended period. Although some published studies have reported positive results following an at-home period using the technology, others have reported little or no change in performance.^{14,38,43–45} The differences may arise from these studies’ lack of controlled/quantified usage at home. 

Selection Bias

We included participants of various ages, diagnoses, and visual acuities, typical of patients in our low-vision service who might benefit from using AT. The clinical presentations of the participants in our study were similar to those of other studies testing AT devices in PVL.^{14,38,45–47} There were no dropouts during the study. Participants received reimbursement for their participation. However, there remains the possibility of volunteer bias in our study.⁴⁸ 

Implications for Clinical Practice and Future Research

We intend for vision professionals to use our findings to guide technology recommendations based on PVL clinical characteristics and functional needs. For example, two options for PVL who retain some pattern vision and wish to access text are image enhancement (e.g. magnifiers) and AI. Our past study on head-mounted displays found that an average of 51% of participants with acquired vision loss and visual acuity better than 20/800 who could not complete the tasks at baseline could complete reading tasks using magnification.²⁴ In our current study, 73% of the participants within this visual acuity range who could not complete the tasks at baseline could complete the text tasks when using AI. Although not directly statistically comparable, these findings suggest that AI could be recommended as an acceptable alternative to magnification in this group of PVL. This is especially the case given the differences in cost and the observation that 97% of people with visual impairment already own a smartphone.⁴⁹ It is logical that PVL with light perception or no light perception would not benefit from using image enhancement approaches. In contrast, an average of 76% of PVL in these lower visual acuity categories who could not complete the text tasks at baseline completed them when using AI. In this case, a strong recommendation for the use of AI is supported. 

Furthermore, PVL wish to find assistance with non-text tasks. Based on our current and past findings, PVL with acuity better than 20/800 benefited less from using image enhancement technology or AI in tasks such as those in our searching and identifying category. In contrast, 67% of those with light perception (LP) or no light perception (NLP) who could not complete the tasks at baseline completed them when using an AAII. However, the recommendations for the use of AI here depend upon the AAII, with 90% of those unable at baseline gaining the ability to complete the searching and identifying tasks when using Seeing AI, 71% when using Envision, but only 37% when using OrCam, and 27% when using Lookout. 

These examples emphasize the value of evidence-based data in guiding recommendations for low-vision technologies. 

Acknowledgments

This work was partly supported by a grant from the American Macular Degeneration Foundation. 

Disclosure: W. Seiple, Astellas Pharmaceuticals, Inc. (C); H.P.A. van de Aa, None; F. Garcia-Pina, None; I. Greco, None; C. Roberts, None; R. van Nispen, Janssen-Cilag NV (C)